Efficiently Detecting DDoS in Heterogeneous Networks: A Parameter-Compressed Vertical Federated Learning approach

The cloud-edge collaborative network owns heterogeneity. Generally, detecting DDoS attacks on heterogeneous networks by data aggregation can greatly improve network service quality. However, recent approaches still face significant communication overhead and data loss. This presents challenges in incrementally managing changes in data features due to the dynamic nature of heterogeneous networks. To address these issues, this paper proposes a parameter-compressed vertical federated learning approach. Our approach leverages the data features of multiple networks while ensuring that only embedding vectors and gradients are exchanged, preventing the training data from leaving the local network. Additionally, we introduce a sparse representation enhancement for the embedding and gradient vectors to minimize communication costs and training time. Furthermore, we propose a feature adaptation scheme for updating the feature dimensions, addressing the need for dynamic network structure adjustments. Experimental results demonstrate that our method effectively reduces communication data size while maintaining high accuracy.